EE 435 Lecture 36 Quantization Noise ENOB Absolute and elative Accuracy DAC Design The String DAC
. eview from last lecture. Quantization Noise in ADC ecall: If the random variable f is uniformly distributed in the interval [A,B] f : U[A,B] then the mean and standard deviation of f are given by A+B B-A μ f = σ= f 2 12 Theorem: If n(t) is a random process, then for large T, t 1+T 2 2 2 1 V = n t dt = σ +μ 1 MS n n T t
. eview from last lecture. Quantization Noise in ADC How does the SN change if the input is a sinusoid that goes from 0 to X EF centered at X EF /2? 0.5X LSB ε Q -0.5X LSB X V LSB MS = 12 But XEF 1 V INMS= 2 2 Thus obtain XEF 2 2 n 3 SN = = 2 XLSB 2 12 Finally, in db, n 3 SN db = 20log 2 =6.02 n + 1.76 2 t
. eview from last lecture. ENOB based upon Quantization Noise SN = 6.02 n + 1.76 Solving for n, obtain ENOB = SNdB-1.76 6.02 Note: could have used the SN db for a triangle input and would have obtained the expression SN db ENOB = 6.02 But the earlier expression is more widely used when specifying the ENOB based upon the noise level present in a data converter
Quantization Noise Effects of quantization noise can be very significant, even at high resolution, when signals are not of maximum magnitude X EF X IN t X EF X IN t Quantization noise remains constant but signal level is reduced The desire to use a data converter at a small fraction of full range Is one of the major reasons high resolution is required
Quantization Noise Effects of quantization noise can be very significant, even at high resolution, when signals are not of maximum magnitude X EF X IN t
Quantization Noise Example: If a 14-bit audio output is derived from a DAC designed for providing an output of 100W but the normal listening level is at 50mW, what is the SN due to quantization noise at maximum output and at the normal listening level? What is the ENOB of the audio system when operating at 50mW? At 100W output, SN=6.02n+1.76 = 90.6dB 2 V =100W L V 2 1 L =50mW V V= 1 44.7 20log 10 V 1 =20log 10 V-20log 10 44.7=20log 10 V -33dB At 50mW output, SN reduced by 33dB to 57.6dB SNdB-1.76 57.6-1.76 ENOB = = = -9.3 6.02 6.02 Note the dramatic reduction in the effective resolution of the DAC when operated at only a small fraction of full-scale.
ENOB Summary esolution: INL: ENOB = log N log 2 10 ACT 10 2 ENOB = n -log -1 log N 2 ACT n specified res, ν INL in LSB DNL: VMAX VMIN ENOB log2 1 MAX V MAX and V MIN are max and min outputs and Δ MAX is maximum absolute step (HW problem) Quantization noise: ENOB = ENOB = SN db 6.02 SNdB-1.76 6.02 rel to triangle/sawtooth rel to sinusoid
Performance Characterization of Data Converters Static characteristics esolution Least Significant Bit (LSB) Offset and Gain Errors Absolute Accuracy elative Accuracy Integral Nonlinearity (INL) Differential Nonlinearity (DNL) Monotonicity (DAC) Missing Codes (ADC) Quantization Noise Low-f Spurious Free Dynamic ange (SFD) Low-f Total Harmonic Distortion (THD) Effective Number of Bits (ENOB) Power Dissipation
Absolute Accuracy Absolute Accuracy is the difference between the actual output and the ideal or desired output of a data converter The ideal or desired output is in reference to an absolute standard (often maintained by the National Bureau of Standards) and could be volts, amps, time, weight, distance, or one of a large number of other physical quantities) Absolute accuracy provides no tolerance to offset errors, gain errors, nonlinearity errors, quantization errors, or noise In many applications, absolute accuracy is not of a major concern but scales, meters, etc. may be more concerned about Absolute accuracy than any other parameter
elative Accuracy In the context of data converters, pseudo-static elative Accuracy is the difference between the actual output and an appropriate fit-line to overall output of the data converter INL is often used as a measure of the relative accuracy In many, if not most, applications, relative accuracy is of much more concern than absolute accuracy Some architectures with good relative accuracy will have very small deviations in the outputs for closely-spaced inputs whereas others may have relatively large deviations in outputs for closely-spaced inputs DNL provides some measure of how outputs for closely-spaced inputs compare X OUT X OUT X EF X EF X IN X IN X EF X EF
DAC Architectures (Nyquist ate) Types Voltage Scaling esistor String DACs (string DACs) Interpolating Current Steering Binarily Weighted esistors -2 Ladders Current Source Steering Thermometer Coded Binary Weighted Segmented Charge edistribution Switched Capacitor Serial Algorithmic Cyclic or e-circulating Pipelined Integrating esistor Switching MDACs (multiplying DACs)
DAC Architectures Structures Hybrid or Segmented Mode of Operation Current Mode Voltage Mode Charge Mode Self-Calibrating Analog Calibration Foreground Background Digital Calibration Foreground Background Dynamic Element Matching Laser of Link Trimmed Thermometer Coded or Binary adix 2 or non-radix 2 Inherently Monotone
DAC Architectures Type of Classification may not be unique nor mutually exclusive Structure is not mutually exclusive All approaches listed are used (and probably some others as well) Some are much more popular than others Popular Architectures esistor String (interpolating) Current Source Steering (with segmentation) Many new architectures are possible and some may be much better than the best currently available All have perfect performance if parasitic and matching performance are ignored! Major challenge is in determining appropriate architecture and managing the parasitics
Nonideal Effects of Concern Matching Parasitic Capacitances (including Charge injection) Loading Nonlinearities Interconnect resistors Noise Slow and plagued by jitter Temperature Effects Aging Package stress
Observations Yield Loss is the major penalty for not appropriately managing parasitics and matching and this loss can be ruthless The ultimate performance limit of essentially all DACs is the yield loss associated with parasitics and matching Many designers do not have or use good statistical models that accurately predict data converter performance If you work of a company that does not have good statistical device models Convince model groups of the importantc of dedveloping these models (or) develop appropriate test strutures to characterize your process Existing nonlinear device models may not sufficiently accurately predict device nonlinearities for high-end data converter applications
Observations Experienced Designers/Companies often produce superior data converter products Essentially all companies have access to the same literature, regularly reverse engineer successful competitors products and key benefits in successful competitors products are generally not locked up in patents High-end designs( speed and resolution) may get attention in the peer community but practical moderate performance converters usually make the cash flow Area (from a silicon cost viewpoint) is usually not the driving factor in high-end designs where attractive price/mfg cost ratios
Data Converter Design Strategies There are many different DAC and ADC architectures that have been proposed and that are in widespread use today Almost all work perfectly if all components are ideal Most data converter design work involves identifying the contributors to nonideal performance and finding work-arounds to these problems Some architectures are more difficult to find work-arounds than others All contributors to nonidealities that are problematic at a given resolution of speed level must be identified and mitigated The effects of not identifying nonidealities generally fall into one of two categories Matching-critical nonidealities (degrade yield) Component nonlinearities (degrade performance even if desired matching is present)
Identifying Problems/Challenges and Clever/Viable Solutions Many problems occur repeatedly so should recognize what they are Identify clever solutions to basic problems they often are useful in many applications Don t make the same mistake twice! The problem: The perceived solution: The practical or clever solution: The List Keeper!
-String DAC V FF X IN n S 1 S 2 V OUT S N-2 S N-1 S N Basic -String DAC
-String DAC V FF S 1 2 n X IN n Binary to Thermometer Decoder S 2 V OUT S N-2 S N-1 S N Basic -String DAC including Logic to Control Switches
-String DAC If all components are ideal, performance of the -string DAC is that of an ideal DAC! V FF 2 n X IN n Key Properties of -String DAC S 1 Binary to Thermometer Decoder One of the simplest DAC architectures -string DAC is inherently monotone S 2 Possible Limitations or Challenges S N-2 V OUT S N-1 S N
-String DAC If all components are ideal, performance of the -string DAC is that of an ideal DAC! V FF 2 n X IN n Key Properties of -String DAC S 1 Binary to Thermometer Decoder One of the simplest DAC architectures -string DAC is inherently monotone S 2 Possible Limitations or Challenges V OUT Binary to Thermometer Decoder (BTTD) gets large for n large S N-2 S N-1 Logic delays in BTTD may degrade performance S N Matching of the resistors may not be perfect Local random variations Gradient effects How can switches be made?
-String DAC Typical strategy for implementing the switch S k V FF 2 n X IN n S 1 Binary to Thermometer Decoder S 2 V OUT d k S N-2 S N-1 M k S N Switch is an analog MUX Very simple structure Switch array combined with the BTTD forms a 2 n :1 analog MUX
-String DAC -String DAC with MOS switches Possible Limitations: V FF d 1 2 n X IN n M 1 d 2 Binary to Thermometer Decoder M 2 d N-2 V OUT M N-2 d N-1 C L M N-1 d N M N
-String DAC -String DAC with MOS switches Possible Limitations: Switch impedance is not 0 V FF d 1 2 n X IN n Switch may not even turn on at all if V EF is large M 1 d 2 Binary to Thermometer Decoder Switch impedance is input-code dependent M 2 Time constants are input-code dependent Transition times are previous-code dependent C L has 2 n diffusion capacitances so can get very large M N-2 d N-2 d N-1 C L V OUT Mismatch of resistors local random variation gradient effects Decoder can get very large for n large M N-1 d N M N outing of the 2n switch signals can become very long and consume lots of area
Basic -String DAC V FF M 1 d 1 d 2 2 n X IN n Binary to Thermometer Decoder M 2 d N-2 V OUT M N-2 d N-1 C L M N-1 d N M N
-String DAC V EF X IN n Decoder b 3 b 3 b 2 b 2 b 1 b 1 -String V OUT Tree Decoder
End of Lecture 36